Transparency · Open questions
Open questions. Shared research.
These research questions shape the next development phase. Explore the tradeoffs and contribute evidence or relevant prior work.
- Q01
How do you verify work that is not deterministic?
Re-execution works for deterministic packets. Much AI work is not deterministic. Cross-model comparison, structural checks and outcome resolution each cover part of the gap, and none of them cover all of it. This is the central unsolved problem.
- Q02
What stops validators from colluding?
Independent validation is only independent if validators are actually distinct parties with distinct incentives. Identity, selection, and penalty design all have to hold simultaneously. During controlled stages this is not yet solved — it is avoided by operating the validators ourselves, which is not a solution.
- Q03
Is reproduction affordable?
If verifying work costs as much as performing it, the network is at best twice the price of a centralized provider. Job classes have to be priced by verification cost, and some classes may never be economically verifiable.
- Q04
How much can a worker see?
Packet isolation limits exposure but does not eliminate it. Jobs that depend on sensitive inputs may not be appropriate for a shared network at all, and we would rather exclude them than pretend otherwise.
- Q05
Can provenance be trusted if identity is pseudonymous?
Pseudonymous but persistent identity supports attribution over time. It does not prevent a determined party from operating many identities. Sybil resistance without permissioning is unresolved.
- Q06
How do you decentralize without being captured?
Removing central coordination transfers power somewhere. Whether that is to capital, to the largest operators or to a governance process is a design decision we have not finalised.
- Q07
What happens to results that turn out to be wrong?
An accepted result can still be wrong. Braid preserves the record so it can be scored later, but how downstream Cortex context should be revised when a claim is invalidated is still being designed.
Push back
Help advance the research
Relevant prior work and reproducible findings are welcome.